A small Pytorch-like Python library for deep learning, written in C++, with CUDA support.

Here are the library features that are implemented:

• Creation of tensors from numpy arrays
• Basic operations on tensors (addition, multiplication, etc.)
• Basic neural network layers (Linear, ReLU)
• Workable Adam optimizer
• Autograd system for automatic differentiation (for most features)
• Support for GPU acceleration (CUDA) (for most features)

Here are the other features of the project:

• Compilation of the library using CMake, Make and Wheel
• Complete testing of the library using Unittest
• Automatic tests using Github Actions (only CPU features)

• No automatic differentiation for max, min, mean, transpose and reshape operations.
• No GPU support for scalar operations (addition, multiplication, etc.), division between tensors, square root and reshape.

• Fill in the missing features
• Add more neural networks layers (Convolution, Sigmoid, etc.)
• Add more optimizers (SGD, etc.)
• Add Loss functions (MSE, CrossEntropy, etc.)
• Add more tests
• Add examples to understand how the library works
• Add documentation
• Deploy the library on PyPI using cibuildwheel

• Lucas de Lima Nogueira for his amazing repo PyNorch and his great tutorial.
• Arthur Dujardin for his amazing repo Nets.